Absa Group, a South African financial services company with multiple banks across Africa, faced a customer service challenge when it split off from Barclays in 2018.
“Because we’ve gone through a lot of changes from Barclays to Absa, we had clear gaps in our call center infrastructure,” said Johan Viljoen, head of channel design at the Africa Regional Office for Absa. “Improving on those required a huge investment."
Absa settled on a virtual assistant in 2019 created by New York-based
Being sensitive to an array of customer needs is important for any bank rolling out a virtual assistant. A chatbot misfire can lead to embarrassment and lost customers who are sent on an endless loop to “please rephrase the question.”
In the U.S., training virtual assistants on limited sets of data can mean simple missteps such as greeting customers in the Northeast with the collective pronoun of “y’all” that would be more familiar to residents of the South. But skewed data can also mean a virtual assistant misinterprets a customer request because of the terminology they use, or addresses underbanked customers in a manner that is more appropriate for high-net-worth individuals, if the assistant was trained to answer questions from wealthy customers.
Sherry Comes, a managing director with Deloitte Consulting, where she leads the conversational AI practice, says there are several ways that banks can train their conversational AI assistants, from using crowdsourcing to analyzing transcripts from live agent calls.
Comes says it’s imperative to consider different nationalities, socioeconomic levels and more when gathering data sets to prevent bias from creeping in.
“The key is diversity,” she said.
Kasisto, which created its Kai virtual assistant for consumer banking, business banking and investment management, has its own methods for de-biasing training data. It collects data from client conversations around the world, looking across demographics, income levels and ethnicities, and has human annotators tag the data that is then used to train the virtual assistant.
Zor Gorelov, Kasisto's CEO, watches out for several ways that data can become skewed. If an assistant is trained to answer questions from a select group, for instance high-net-worth customers, it may not interact as well with customers who are underbanked. People interact with live chat and call center agents differently than they do with virtual assistants, making that a poor source of training data as well; for example, customers are unlikely to call an agent and request a list of recent transactions, as they may with a virtual assistant.
Relying on data from customers in one country could mangle a virtual assistant’s responses for customers of a division of that bank in another country if they use different terms to request the same action.
Kasisto’s virtual assistant is deployed in 16 countries and converses in four languages. “Our approach has always been to aggregate and anonymize data from banks around the world and rationalize it across various demographics and income levels,” said Gorelov.
At first, Absa installed a version of Abby that answered frequently asked questions. In 2020, it ramped up Abby’s capabilities to handle authenticated transactions as well, such as paying bills or checking balances. Customers can access Abby through their country’s public-facing Absa website or get instant replies (and log in to their accounts to execute transactions) on the messaging app WhatsApp.
The timing of its launch “was perfect,” Viljoen said.
“As we moved from Barclays to Absa and introduced the bot across multiple markets, COVID was introduced into the equation, and we’ve been able to support the call centers and alleviate a lot of the congestion,” he said.
Viljoen and his team started with a standard set of questions that Kasisto provided. In some cases, financial products were similar enough that they could apply the same response across the board. In other instances, they tweaked responses for different countries.
They have had to adapt to subtle differences across their footprint so the interactions with Abby flow smoothly. For example, customers in some countries who want to pay a bill might reference the type of bill by requesting “pay utility” rather than the more generic “pay a bill.”
Tone matters too.
“We’ve gone for a mild approach with Abby’s tone,” said Viljoen. “Being witty can be misconstrued in some of our countries.”
Human oversight is critical as well. Absa has an “annotation lead” in each of its countries. These specialists lead a team of annotators who parse through conversational AI data to ensure Abby’s replies align with what the customers are asking, as well as detect trends in the questions and answers that may influence future topics or how Absa markets its products.
“We see annotators as the voice of the customer,” Viljoen said.
Some of Kasisto’s clients perform data annotation on their own, while others wait for Kasisto to deliver natural language understanding updates as part of regular software releases. The question and utterance data from Kasisto clients around the world are reviewed on a regular basis and are used to improve NLU models within each release.
The next step for Absa and Kasisto is teaching Abby languages that are the primary languages in some of Absa’s countries, namely Portuguese and Swahili.
For now, customers can have fun bantering with Abby in English.
Ask the WhatsApp version of Abby in Botswana what movies it likes and it responds, “I never have time for movies, I am always online and ready to serve you.” As for keeping itself entertained, “I talk to you for fun,” replies Abby. “I know this might sound nerdy, but I learn for fun as well.”
